New Breakthrough in Active Domain Adaptation: Uncertainty-Guided Region Coverage

Active Domain Adaptation (ADA) aims to reduce annotation costs in both source and target domains through active learning strategies while improving the performance of unsupervised domain adaptation. However, traditional methods typically optimize for average accuracy, which can lead to performance degradation on certain difficult categories and affect model robustness in real-world scenarios. To address this issue, a recent study proposes an uncertainty-guided region coverage method that surpasses existing baselines on multiple domain adaptation benchmarks.

Limitations of Point-wise Selection and Region-based Improvements

Early research attempted to use influence functions for point-wise sample selection to improve domain adaptation performance. But the study points out that point-wise selection strategies suffer from sensitivity and redundancy issues, easily leading to overfitting of the acquisition function. Specifically, influence estimates for individual samples may be affected by noise, and repeatedly selecting similar samples reduces annotation efficiency. To this end, the authors shift to a region-based approach, ensuring improvement of the adapted model by maximizing positive manifold coverage. This region coverage strategy not only reduces redundancy but also more stably enhances the model's generalization ability on the target domain.

Foundations of Domain Adaptation and Active Learning: Domain Adaptation addresses the problem of inconsistent data distributions between source and target domains. Unsupervised Domain Adaptation (UDA) assumes the target domain has no labels, transferring knowledge through feature alignment or adversarial training. Active Domain Adaptation (ADA) allows querying labels for a small number of target domain samples, trading annotation cost for performance gains. The core of active learning is the acquisition function, which selects the most informative samples based on criteria such as uncertainty and diversity. Traditional ADA mostly optimizes average accuracy, ignoring inter-class differences, leading to performance collapse on difficult categories (e.g., classes with few samples or large divergence from the source domain). Category-Aware ADA was thus proposed to balance performance across categories, with the challenge being how to simultaneously improve overall and minority-class accuracy under a limited annotation budget.

Uncertainty-Guided Decision Boundary Optimization

Building on region coverage, the research further introduces an uncertainty score into the selection mechanism. This score guides the model to learn better decision boundaries on the target domain. By prioritizing samples in high-uncertainty regions, the model can more effectively correct classification errors, especially in regions with ambiguous class boundaries. Experiments show that this uncertainty-guided strategy significantly improves recognition accuracy on difficult categories, not just overall average accuracy.

Influence Functions and Manifold Coverage: The Influence Function originates from robust statistics and is used to estimate the degree of influence a single training sample has on model parameters or predictions. In active learning, it can approximate the change in target domain loss after a sample is annotated, thereby guiding selection. However, point-wise influence estimation is sensitive to noise, and highly similar samples have close influence values, leading to repeated selection. Positive manifold coverage, from a geometric perspective, treats sample selection as maximizing the coverage area on the target domain feature manifold, ensuring that selected samples can represent the entire data distribution and reducing redundancy. Uncertainty scores are often based on prediction entropy, confidence, or Bayesian approximation. High-uncertainty samples tend to lie near decision boundaries, and annotating them can most effectively refine the boundary. The two are combined: region coverage ensures diversity, uncertainty ensures informativeness, jointly optimizing category-aware active domain adaptation.

Experimental Validation and Performance Comparison

The authors conducted comprehensive experiments on multiple domain adaptation benchmarks, comparing the proposed method with the first category-aware active domain adaptation baseline and existing active domain adaptation methods. Results show that the method leads in both overall accuracy and class balance. You may not have noticed that the combination of region coverage and uncertainty scoring not only alleviates the category performance degradation problem but also reduces annotation costs. The code has been open-sourced on GitHub for the community to reproduce and conduct further research.

Benchmarks and Evaluation Metrics: Common domain adaptation benchmarks include Office-31, Office-Home, VisDA-2017, and DomainNet, covering different domain shifts and numbers of categories. Evaluation considers not only overall accuracy but also per-class average accuracy and worst-class accuracy to measure class balance. The first category-aware active domain adaptation baseline typically refers to CDUAL or similar methods, which introduce class-level uncertainty or fairness constraints. Annotation cost is measured by the number of queried samples or budget ratio, and active domain adaptation must be compared under the same budget. Experiments typically employ multiple rounds of active learning cycles: train model, select samples, query labels, retrain. Open-source code (e.g., GitHub repositories) provides reproduction details, including network architecture, optimizer, and hyperparameters, promoting fair comparison.

Implications for Active Domain Adaptation

The core contribution of this work lies in extending sample selection from point-wise to region-based and incorporating uncertainty measures, providing new ideas for category-aware active domain adaptation. In the future, this method is expected to be applied to more complex real-world scenarios, such as autonomous driving and medical image analysis, where accurate recognition of difficult categories is crucial. Meanwhile, the generality of the region coverage strategy may also inspire research in other active learning tasks.

One-line Summary

Uncertainty-guided region coverage method improves category-aware active domain adaptation performance, balancing overall and difficult category accuracy.

Paragraph Summary

This study addresses the problem of performance degradation on difficult categories caused by average accuracy optimization in active domain adaptation by proposing an uncertainty-guided region coverage method. Early point-wise selection suffers from sensitivity and redundancy, easily overfitting the acquisition function; the authors instead adopt a region-based strategy, reducing redundancy by maximizing positive manifold coverage and introducing uncertainty scores to optimize target domain decision boundaries. On multiple domain adaptation benchmarks, the method surpasses existing baselines, including the first category-aware active domain adaptation method, in both overall accuracy and class balance, while reducing annotation costs. The code has been open-sourced, providing a reproducible foundation for future research.